HistoriQA: New Multi-Hop QA Dataset for French History.

Historical research often involves complex reasoning, requiring the synthesis of information from multiple, heterogeneous sources and an understanding of temporal relationships. Traditional NLP benchmarks frequently fall short in capturing these nuanced patterns, creating a gap between general language model capabilities and the specific needs of historical scholarship. To address this, researchers have introduced HistoriQA-ThirdRepublic, a novel French-language dataset specifically designed for multi-hop historical question answering. This corpus is built from parliamentary debates and newspapers spanning the French Third Republic (1870-1940) and was developed in close collaboration with a historian to ensure its relevance and accuracy. The dataset comprises 1782 questions that emphasize multi-hop connections across diverse historical documents, demanding cross-source synthesis, temporal reasoning, and the integration of sparse evidence. HistoriQA-ThirdRepublic serves as a valuable resource for evaluating the performance of retrieval-augmented and large language models in domain-specific contexts, demonstrating how the methodology can be adapted for other languages and national corpora, thereby bridging the divide between advanced NLP and the practical demands of historical inquiry.